黄月平, 李小锋, 齐乃新, 卢瑞涛, 张胜修. 基于难例挖掘和自适应时间正则化的视觉目标跟踪算法[J]. 机器人, 2021, 43(3): 350-363. DOI: 10.13973/j.cnki.robot.200370
引用本文: 黄月平, 李小锋, 齐乃新, 卢瑞涛, 张胜修. 基于难例挖掘和自适应时间正则化的视觉目标跟踪算法[J]. 机器人, 2021, 43(3): 350-363. DOI: 10.13973/j.cnki.robot.200370
HUANG Yueping, LI Xiaofeng, QI Naixin, LU Ruitao, ZHANG Shengxiu. Visual Object Tracking Algorithm Based on Hard Negative Mining and Adaptive Temporal Regularization[J]. ROBOT, 2021, 43(3): 350-363. DOI: 10.13973/j.cnki.robot.200370
Citation: HUANG Yueping, LI Xiaofeng, QI Naixin, LU Ruitao, ZHANG Shengxiu. Visual Object Tracking Algorithm Based on Hard Negative Mining and Adaptive Temporal Regularization[J]. ROBOT, 2021, 43(3): 350-363. DOI: 10.13973/j.cnki.robot.200370

基于难例挖掘和自适应时间正则化的视觉目标跟踪算法

Visual Object Tracking Algorithm Based on Hard Negative Mining and Adaptive Temporal Regularization

  • 摘要: 针对复杂情况下视觉目标跟踪算法性能严重退化的问题,提出一种基于难例挖掘和自适应时间正则化的视觉目标跟踪算法.首先,该算法在Staple算法基础上,深度挖掘困难负样本用于相关滤波器训练,提高了跟踪算法的抗干扰能力;其次,加入自适应时间正则化约束,根据目标响应图的变化情况,自适应确定时间正则化系数及模型更新策略,增强了跟踪算法的鉴别能力.在数据集OTB-2015、TC-128和UAV123上的实验结果表明,该算法能够有效应对复杂情况下跟踪性能退化的问题,且运行速度超过30帧/秒,满足实时性需求,综合性能优于对比算法.

     

    Abstract: To alleviate the severe performance degradation of visual object tracking algorithms in complex situations, a visual object tracking algorithm based on hard negative mining and adaptive temporal regularization is proposed. Firstly, hard negative samples are deeply mined based on the Staple algorithm to train the correlation filter, which improves the anti-jamming ability of the tracking algorithm. Secondly, the adaptive temporal regularization term is added, and the coefficient of temporal regularization term and the model updating strategy are adaptively determined by the response map variation, which improves the identification ability of the tracking algorithm. The experimental results on datasets OTB-2015, TC-128 and UAV123 show that the proposed algorithm can effectively deal with the tracking performance degradation in complex situations, and its running speed is more than 30 frames per second, which meets the real-time requirement. Its comprehensive performance is better than the comparative algorithms.

     

/

返回文章
返回